lists.openwall.net   lists  /  announce  owl-users  owl-dev  john-users  john-dev  passwdqc-users  yescrypt  popa3d-users  /  oss-security  kernel-hardening  musl  sabotage  tlsify  passwords  /  crypt-dev  xvendor  /  Bugtraq  Full-Disclosure  linux-kernel  linux-netdev  linux-ext4  linux-hardening  linux-cve-announce  PHC 
Open Source and information security mailing list archives
 
Hash Suite: Windows password security audit tool. GUI, reports in PDF.
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <3fc11402-53e1-4325-a3ee-5ebd616b5b63@amazon.co.uk>
Date: Tue, 6 Aug 2024 16:39:24 +0100
From: Patrick Roy <roypat@...zon.co.uk>
To: Elliot Berman <quic_eberman@...cinc.com>, Andrew Morton
	<akpm@...ux-foundation.org>, Paolo Bonzini <pbonzini@...hat.com>, "Sean
 Christopherson" <seanjc@...gle.com>, Fuad Tabba <tabba@...gle.com>, "David
 Hildenbrand" <david@...hat.com>, <qperret@...gle.com>, Ackerley Tng
	<ackerleytng@...gle.com>
CC: <linux-coco@...ts.linux.dev>, <linux-arm-msm@...r.kernel.org>,
	<linux-kernel@...r.kernel.org>, <linux-mm@...ck.org>, <kvm@...r.kernel.org>,
	James Gowans <jgowans@...zon.com>, "Kalyazin, Nikita"
	<kalyazin@...zon.co.uk>, "Manwaring, Derek" <derekmn@...zon.com>, "Cali,
 Marco" <xmarcalx@...zon.co.uk>
Subject: Re: [PATCH RFC 3/4] mm: guest_memfd: Add option to remove guest
 private memory from direct map


Hi Elliot,

On Mon, 2024-08-05 at 19:34 +0100, Elliot Berman wrote:
> This patch was reworked from Patrick's patch:
> https://lore.kernel.org/all/20240709132041.3625501-6-roypat@amazon.co.uk/

yaay :D

> While guest_memfd is not available to be mapped by userspace, it is
> still accessible through the kernel's direct map. This means that in
> scenarios where guest-private memory is not hardware protected, it can
> be speculatively read and its contents potentially leaked through
> hardware side-channels. Removing guest-private memory from the direct
> map, thus mitigates a large class of speculative execution issues
> [1, Table 1].
> 
> Direct map removal do not reuse the `.prepare` machinery, since
> `prepare` can be called multiple time, and it is the responsibility of
> the preparation routine to not "prepare" the same folio twice [2]. Thus,
> instead explicitly check if `filemap_grab_folio` allocated a new folio,
> and remove the returned folio from the direct map only if this was the
> case.

My patch did this, but you separated the PG_uptodate logic from the
direct map removal, right?

> The patch uses release_folio instead of free_folio to reinsert pages
> back into the direct map as by the time free_folio is called,
> folio->mapping can already be NULL. This means that a call to
> folio_inode inside free_folio might deference a NULL pointer, leaving no
> way to access the inode which stores the flags that allow determining
> whether the page was removed from the direct map in the first place.

I thought release_folio was only called for folios with PG_private=1?
You choose PG_private=1 to mean "this folio is in the direct map", so it
gets called for exactly the wrong folios (more on that below, too).

> [1]: https://download.vusec.net/papers/quarantine_raid23.pdf
> 
> Cc: Patrick Roy <roypat@...zon.co.uk>
> Signed-off-by: Elliot Berman <quic_eberman@...cinc.com>
> ---
>  include/linux/guest_memfd.h |  8 ++++++
>  mm/guest_memfd.c            | 65 ++++++++++++++++++++++++++++++++++++++++++++-
>  2 files changed, 72 insertions(+), 1 deletion(-)
> 
> diff --git a/include/linux/guest_memfd.h b/include/linux/guest_memfd.h
> index be56d9d53067..f9e4a27aed67 100644
> --- a/include/linux/guest_memfd.h
> +++ b/include/linux/guest_memfd.h
> @@ -25,6 +25,14 @@ struct guest_memfd_operations {
>         int (*release)(struct inode *inode);
>  };
> 
> +/**
> + * @GUEST_MEMFD_FLAG_NO_DIRECT_MAP: When making folios inaccessible by host, also
> + *                                  remove them from the kernel's direct map.
> + */
> +enum {
> +       GUEST_MEMFD_FLAG_NO_DIRECT_MAP          = BIT(0),
> +};
> +
>  /**
>   * @GUEST_MEMFD_GRAB_UPTODATE: Ensure pages are zeroed/up to date.
>   *                             If trusted hyp will do it, can ommit this flag
> diff --git a/mm/guest_memfd.c b/mm/guest_memfd.c
> index 580138b0f9d4..e9d8cab72b28 100644
> --- a/mm/guest_memfd.c
> +++ b/mm/guest_memfd.c
> @@ -7,9 +7,55 @@
>  #include <linux/falloc.h>
>  #include <linux/guest_memfd.h>
>  #include <linux/pagemap.h>
> +#include <linux/set_memory.h>
> +
> +static inline int guest_memfd_folio_private(struct folio *folio)
> +{
> +       unsigned long nr_pages = folio_nr_pages(folio);
> +       unsigned long i;
> +       int r;
> +
> +       for (i = 0; i < nr_pages; i++) {
> +               struct page *page = folio_page(folio, i);
> +
> +               r = set_direct_map_invalid_noflush(page);
> +               if (r < 0)
> +                       goto out_remap;
> +       }
> +
> +       folio_set_private(folio);

Mh, you've inverted the semantics of PG_private in the context of gmem
here, compared to my patch. For me, PG_private=1 meant "this folio is
back in the direct map". For you it means "this folio is removed from
the direct map". 

Could you elaborate on why you require these different semantics for
PG_private? Actually, I think in this patch series, you could just drop
the PG_private stuff altogether, as the only place you do
folio_test_private is in guest_memfd_clear_private, but iirc calling
set_direct_map_default_noflush on a page that's already in the direct
map is a NOOP anyway.

On the other hand, as Paolo pointed out in my patches [1], just using a
page flag to track direct map presence for gmem is not enough. We
actually need to keep a refcount in folio->private to keep track of how
many different actors request a folio's direct map presence (in the
specific case in my patch series, it was different pfn_to_gfn_caches for
the kvm-clock structures of different vcpus, which the guest can place
into the same gfn). While this might not be a concern for the the
pKVM/Gunyah case, where the guest dictates memory state, it's required
for the non-CoCo case where KVM/userspace can set arbitrary guest gfns
to shared if it needs/wants to access them for whatever reason. So for
this we'd need to have PG_private=1 mean "direct map entry restored" (as
if PG_private=0, there is no folio->private).

[1]: https://lore.kernel.org/kvm/20240709132041.3625501-1-roypat@amazon.co.uk/T/#m0608c4b6a069b3953d7ee97f48577d32688a3315

> +       return 0;
> +out_remap:
> +       for (; i > 0; i--) {
> +               struct page *page = folio_page(folio, i - 1);
> +
> +               BUG_ON(set_direct_map_default_noflush(page));
> +       }
> +       return r;
> +}
> +
> +static inline void guest_memfd_folio_clear_private(struct folio *folio)
> +{
> +       unsigned long start = (unsigned long)folio_address(folio);
> +       unsigned long nr = folio_nr_pages(folio);
> +       unsigned long i;
> +
> +       if (!folio_test_private(folio))
> +               return;
> +
> +       for (i = 0; i < nr; i++) {
> +               struct page *page = folio_page(folio, i);
> +
> +               BUG_ON(set_direct_map_default_noflush(page));
> +       }
> +       flush_tlb_kernel_range(start, start + folio_size(folio));
> +
> +       folio_clear_private(folio);
> +}
> 
>  struct folio *guest_memfd_grab_folio(struct file *file, pgoff_t index, u32 flags)
>  {
> +       unsigned long gmem_flags = (unsigned long)file->private_data;
>         struct inode *inode = file_inode(file);
>         struct guest_memfd_operations *ops = inode->i_private;
>         struct folio *folio;
> @@ -43,6 +89,12 @@ struct folio *guest_memfd_grab_folio(struct file *file, pgoff_t index, u32 flags
>                         goto out_err;
>         }
> 
> +       if (gmem_flags & GUEST_MEMFD_FLAG_NO_DIRECT_MAP) {
> +               r = guest_memfd_folio_private(folio);
> +               if (r)
> +                       goto out_err;
> +       }
> +

How does a caller of guest_memfd_grab_folio know whether a folio needs
to be removed from the direct map? E.g. how can a caller know ahead of
time whether guest_memfd_grab_folio will return a freshly allocated
folio (which thus needs to be removed from the direct map), vs a folio
that already exists and has been removed from the direct map (probably
fine to remove from direct map again), vs a folio that already exists
and is currently re-inserted into the direct map for whatever reason
(must not remove these from the direct map, as other parts of
KVM/userspace probably don't expect the direct map entries to disappear
from underneath them). I couldn't figure this one out for my series,
which is why I went with hooking into the PG_uptodate logic to always
remove direct map entries on freshly allocated folios.

>         /*
>          * Ignore accessed, referenced, and dirty flags.  The memory is
>          * unevictable and there is no storage to write back to.
> @@ -213,14 +265,25 @@ static bool gmem_release_folio(struct folio *folio, gfp_t gfp)
>         if (ops->invalidate_end)
>                 ops->invalidate_end(inode, offset, nr);
> 
> +       guest_memfd_folio_clear_private(folio);
> +
>         return true;
>  }
> 
> +static void gmem_invalidate_folio(struct folio *folio, size_t offset, size_t len)
> +{
> +       /* not yet supported */
> +       BUG_ON(offset || len != folio_size(folio));
> +
> +       BUG_ON(!gmem_release_folio(folio, 0));
> +}
> +
>  static const struct address_space_operations gmem_aops = {
>         .dirty_folio = noop_dirty_folio,
>         .migrate_folio = gmem_migrate_folio,
>         .error_remove_folio = gmem_error_folio,
>         .release_folio = gmem_release_folio,
> +       .invalidate_folio = gmem_invalidate_folio,
>  };
> 
>  static inline bool guest_memfd_check_ops(const struct guest_memfd_operations *ops)
> @@ -241,7 +304,7 @@ struct file *guest_memfd_alloc(const char *name,
>         if (!guest_memfd_check_ops(ops))
>                 return ERR_PTR(-EINVAL);
> 
> -       if (flags)
> +       if (flags & ~GUEST_MEMFD_FLAG_NO_DIRECT_MAP)
>                 return ERR_PTR(-EINVAL);
> 
>         /*
> 
> --
> 2.34.1
> 

Best, 
Patrick

Powered by blists - more mailing lists

Powered by Openwall GNU/*/Linux Powered by OpenVZ